Energy consumption of the Internet is already substantial and it is likely to increase as operators deploy faster equipment to handle popular bandwidth-intensive services, such as streaming and video-on-demand. Further, the CMOS technology that is predominantly used in the ICT equipment is reaching its energy efficiency limits. The net effect of these issues is that the Internet will not be able to continue its exponential growth without a substantial increase in the energy used for its power and cooling. We are designing energy-proportional networks that try to use the minimum amount of energy to satisfy each offered load (traffic).
The biggest fraction of wired network energy is consumed by the access network, even though the individual access devices themselves have small power requirements. Unfortunately, straightforward techniques for sleeping during idle periods are inefficient in this environment because of the continuous lightweight traffic, and the absence of alternative paths that could carry the traffic. We propose two simple techniques: 1) wireless user traffic aggregation that enables access devices to firmly sleep, and 2) switching at the ISP side that significantly increases the number of power-hungry line cards that can sleep. Our thorough evaluation using trace-based simulation and a live prototype shows that for typical urban settings it is possible to save 66% of access network energy on average, while being within 7-35% of the computationally-hard optimal aggregation and impractical optimal switching.
We argue that a complete solution for reducing the energy consumption of the Internet core and datacenter networks requires a network-wide approach that works in conjunction with local measures. However, the large body of work on intradomain routing and traffic engineering shows the difficulty in meeting the desired network properties, and adding energy-awareness makes the problem even harder. Our key contribution is in adopting a hybrid approach in which the responsibilities are divided between the offline and online components. Doing so allows us to make best use of each component, and address the practical challenges of achieving energy proportionality in networks: responsiveness, stability, ability to match network resources to the load, and ease of deployment.
N. Vasic, D. Novakovic, S. Miucin, D. Kostic and R. Bianchini. DejaVu: Accelerating Resource Allocation in Virtualized Environments. Proceedings of the Seventeenth International Conference on Architectural Support for Programming Languages and Operating Systems (ASPLOS), 2012.
Vasic, N., Novakovic, D., Shekhar, S., Bhurat, P., Canini, M. and Kostic, D., Identifying and Using Energy-Critical Paths, Proceedings of ACM CoNEXT 2011, December 2011.
Vasic, N., Novakovic, D., Shekhar, S., Bhurat, P., Canini, M. and Kostic, D., Identifying and Using Energy-Critical Paths, EPFL Technical Report EPFL-REPORT-168767, June 2011.
Goma, E., Canini, M., Lopez, A., Laoutaris, N., Kostic, D., Rodriguez, P., Stanojevic, R., and Yague, P., Insomnia in the Access (or How to Curb Access Network Related Energy Consumption), Proceedings of the ACM SIGCOMM 2011 Conference on Applications, Technologies, Architectures, and Protocols for Computer Communications, August 2011.
Vasic, N., Novakovic, D., Shekhar, S., Bhurat, P., Canini, M. and Kostic, D., Responsive, Energy-Proportional Networks, EPFL Technical Report, July 2010.
In the news
- Swiss NSF Horizons newsletter (page 15) , December 2013.
- EPFL news and research highlights, September 2011.
- This research is sponsored by the Swiss National Science Foundation (grant FNS 200021-130265). Nedeljko Vasic was supported by an IBM PhD Fellowship for the 2009/2010 academic year.